learning_world_model.py

class LearningWorldModel:
    """
    Learns environment dynamics from experience.
    """

    def __init__(self):

        # (state, action) -> observed next state samples
        self.memory = []

        # simple learned weights (placeholder model)
        self.weights = {
            "move": -0.05,
            "interact": -0.1,
            "idle": 0.01
        }

    # =====================================================
    # PREDICT
    # =====================================================
    def predict(self, state, action):

        new_state = state.copy()

        delta = self.weights.get(action, 0)

        new_state["energy"] = state.get("energy", 1.0) + delta

        return new_state

    # =====================================================
    # OBSERVE REAL RESULT
    # =====================================================
    def observe(self, state, action, next_state):

        self.memory.append((state, action, next_state))

        if len(self.memory) > 1000:
            self.memory.pop(0)

        self._update_model()

    # =====================================================
    # LEARNING STEP
    # =====================================================
    def _update_model(self):

        # very simplified learning rule
        for state, action, next_state in self.memory[-50:]:

            predicted = self.predict(state, action)

            error = next_state["energy"] - predicted["energy"]

            # gradient-like update
            self.weights[action] = self.weights.get(action, 0) + 0.01 * error